Prior literature provides conflicting evidence about the impact of speculation on gold futures returns, volatility, and the relationship between market fundamentals and prices. In this paper, we exploit trade volume information to determine the most appropriate family of factors to adopt when modelling gold futures. Using the Disaggregated Commitment of Traders report, we find that extreme levels of speculation are informative in that they signify a shift in the relative modelling accuracy of macroeconomic and latent factors. A simple composite prediction framework, incorporating the changing level of speculation, empirically demonstrates the uncovered phenomenon and offers improved predictive accuracy for gold futures prices. Furthermore, our findings are shown to be robust to alternative latent and macroeconomic model specifications.